|
|
Allocation Policies
The purpose of this document is to describe the policies and procedures
that will be followed in allocating time on System X.
- Allocations will be in units of cpu-hours. Hours for internal users
and collaborating researchers are free. External users MUST
contact Terry Herdman (herd88 at vt.edu) or Cal Ribbens (ribbens at vt.edu)
to negotiate for time on System X.
- Each allocation will create a bank account of hours. The accounting
system refers to this bank account as a hat. Each hat will
have a principal investigator (PI) who decides which users are allowed
to run jobs under that hat. A single user may run jobs under more
than one hat, e.g., a user who is part of two research groups.
- Every job submitted to the queuing system must indicate the hat
it is run under. Jobs submitted under hats whose balance has fallen
to zero will not run. No jobs may be run on System X other than through
the queuing system. Users who require a number of nodes to be reserved
exclusively for them for a designated period of time will be charged
as if they used all those nodes for that entire time.
- Each hat will have an end date. Hats will ordinarily last for approximately
one year. Allocations will disappear at their end date. A renewal
process will be established to allow investigators to extend the time
of their original allotment or request additional hours.
- Hats will be assigned priority levels, which will influence the
service levels users experience from the queuing system. For hats
authorized directly by one of the stakeholders or by the Director,
the priority will be set by the authorizer. For hats authorized by
the Allocation Committee, the following guidelines will be used:
- Priority 1. Any hat that is tied to grant money. Plus internally
funded hats that the Allocation Committee identifies as having
the highest scientific merit, i.e., highly scalable code, potentially
high impact, capability computing.
- Priority 2. Reasonable scientific merit. Parallel code exists
but performance and scalability may be unknown at this point.
Capacity computing.
- Priority 3. Exploratory research. Almost a courtesy hat. Not
expecting to make immediate large use of System X, but some preliminary
work is justified.
- Note that the committee may decide to raise the priority
level for a hat, if results justify this.
- Priority level is only one of the factors that determine actual
node assignments by the queuing system. All other things being equal,
a job from a priority Priority 1 hat will run before a job from a
priority Priority 2 hat. However, cluster queuing systems take many
factors into account when deciding which jobs to schedule next. Factors
include: how many jobs this user already has running on the system,
how many nodes are requested and for how long, how many nodes are
currently free on the system. The System X queuing parameters will
be monitored and adjusted to reflect polices described in this document
(e.g., with respect to priority levels), to boost utilization, to
meet the high-level goals of VT ARC, and to maintain a reasonable
degree of fairness among users.
- Requirements for proposals to the Allocation Committee are posted
on the ARC website. Requests for small allocations (up to 10,000 hours)
require much less justification than requests for medium allocations
(up to 100,000 hours), which require less justification than requests
for large allocations (over 100,000 hours).
- Investigators awarded large or medium allocations by the Allocation
Committee will be expected to supply brief summaries of results obtained
on System X.
|
|
